Regarding pluggable iterfaces -Do we really need this? The only place I saw it used it did not seem to matter where the split occured and there was now endchunck/beginchunck stuff. I would MUCH rather see the interface code take care of this with out mucking up the core code (as the eth code already does). Did I miss something here?
The version I have lets a user to choose the interface by supplying appropriate command line. (e.g. kgdbwait kgdb8250=... or kgdbwait kgdbeth=...) It supports an arbitrary number of interfaces. The kgdb core itself is independent of an interface. All interfaces are defined by a structure described below. An interface registers itself with kgdb core by assigning this structure to pointer kgdb_serial.
struct kgdb_serial {
int chunksize;
int (*read_char)(void);
void (*write_char)(int);
void (*flush)(void);
int (*hook)(void);
void (*begin_session)(void);
void (*end_session)(void);
};
Where chunksize is maximum chunksize an interface can handle.
read_char and write_char are derived from getDebugChar and putDebugChar
flush flushes written characters. Flush control is given to kgdb core so that it can ensure that #checksum doesn't split.
begin_session and end_session inform an interface about a gdb communication session. (Haven't decided about console packets to gdb yet)
hook is interface initialization. It can return errors. This allows kgdb core to probe the interface for availability at multiple points. Because of this, there can be multiple debugger entry points
1. At very begining of start_kernel -> Only an 8250 interface with early boot enabled can respond to hook call.
2. After smp initialization -> An 8250 interface without an early boot can respond to this.
3. An ethernet interface can itself call debugger_entry to enter debugger after it's brought up from userland.
Other interfaces can come up at (1) or (2)
On Monday 12 Jan 2004 12:19 pm, Matt Mackall wrote:
On Sun, Jan 11, 2004 at 09:41:57PM -0800, George Anzinger wrote:
For the internal kgdb stuff I have created kdgb_local.h which I intended
to be local to the workings of kgdb and not to contain anything a user
would need.
Agreed, I just haven't touched it since you last mentioned it.
+struct kgdb_hook {
+ char *sendbuf;
+ int maxsend;
I don't see the need of maxsend, or sendbuff, for that matter, as kgdb
uses it now (for the eth code) it is redundant, in that the eth putchar
also does the same thing as is being done in the kgdb_stub.c code. I
think this should be removed from the stub and the limit in the ethcode
relied upon.
Fair enough.
void
putDebugChar(int c)
{
- if (!kgdboe) {
- tty_putDebugChar(c);
- } else {
- eth_putDebugChar(c);
- }
+ if (kh)
+ kh->putchar(c);
}
I was thinking that this might read something like:
if (xxx[kh].putchar(c))
xxx[kh].putchar(c);
One might further want to do something like:
if (!xxx[kh].putchar(c))
kh = 0;
In otherwords, an array (xxx must, of course, be renamed) of stuct
kgdb_hook (which name should also be changed to relate to I/O,
kgdb_IO_hook, for example). Then reserve entry 0 for the rs232 I/O code.
Dunno about that. Probably should work more like the console code,
whoever registers first wins. Early boot will probably be the
exclusive province of serial for a while yet, but designing it in is
probably short-sighted.
An alternate possibility is an array of pointer to struct kgdb_hook
which allows one to define the struct contents as below and to build the
array, all at compile/link time. A legal entry MUST define get and put,
but why not define them all, using dummy functions for the ones that make
no sense in a particular interface.
Throwing all the stubs in a special section could work well too. Then
we could add an avail() function so that early boot debugging could
discover if each one was available. The serial code could use this to
kickstart itself while the eth code could test a local initialized
flag and say "not a chance". Which gives us all the architecture to
throw in other trivial interfaces (parallel, bus-snoopers, etc.).